How do you defend against Liverpool?
We analyse tracking data of 19 Liverpool goals from the latter half of their historic 18–19 campaign, in the form of an Opposition Analysis. We conceptualise a physics-based approach for quantifying the difficulty of a pass called Pass Difficulty, PD
In this post, we won’t look too much into the usual Liverpool passing and shooting stats that we’re already familiar with. The tracking data is a good opportunity to investigate the finer stylistic elements of their game, with Pitch Control.
If you really wanted to defend Liverpool, you’ll also want to learn from Liverpool’s low xG plays. For now, we only have goal data.
We identify 6 points that need addressing, to douse Liverpool’s firepower. We see in these 19 goals, a reiteration of Liverpool’s known strengths: pace, direct passing, and work rate. We propose neutralising these, and forcing Liverpool to rely more on intricate or creative passing, and dribbling.
In this process, we conceptualise a physics-based approach for quantifying the difficulty of a pass called Pass Difficulty, PD. Mainly inspired by Trent Alexander-Arnold’s knack for ridiculous long balls and crosses.
1: Know Who You’re Up Against
These players make things happen out of thin air.
Consider these 2 goals created by Trent Alexander-Arnold. In the first play, Wolverhampton was not disciplined enough in covering the wings, and allowed a cross from a dangerous position. But in the second play, it seems that the Norwich defence didn’t do much wrong. Here, we suggest a Physics-based metric to illustrate Trent’s capability of completing difficult passes. The idea is to quantify which players routinely make difficult passes, and allow them less ball time. Technical discussion on Pass Difficulty, PD in Appendix 1.
For now, we use Expected Threat (xT) as a proxy for Pass Difficulty. Wolves allowed Trent a position that Premier League teams score from 13.6% of the time (probably much higher for Trent!). Against Norwich however, the position Trent crossed from led to a goal in only 2.4% of plays. More on Singh’s xT model.
Now consider this Sadio Mane goal against Bayern. Notice how at the instant Van Dijk passes, the Pitch Control model doesn’t even show any inside leverage for Mane against the Bayern defender. Yet, by the time the ball is mid-air, Mane has a clear patch of red ahead of him. We see this same kind of precision and pace to beat defenders who were in a good position in 4 other goals.
Learnings
- Identify dangerous passers pre-game based on their current form in difficult passes. For now, we use xT, next step is to implement Pass Difficulty (PD).
- When Liverpool build-up from their own half, give Salah and Mane a man coverage ‘cushion’. Bigger cushion for slower defenders. (Like an NFL cornerback.)
2: Don’t get Counter-Attacked
Over half the goals we see were scored following a change of possession. The majority of which were pure counter-attacks, where the ball and opposition started well within the Liverpool half. Watch how the entire pitch goes from blue to red, in a mere 13 second span.
Why are they so effective?
Liverpool employ a very direct style of attacking, with long passes directly into the path of their front-3. This proves to be very effective for counter-attacking as it doesn’t allow any time for the opposition to regain defensive position to make a tackle or force the run wide. Once again, the front 3 show exceptional composure during these fast runs. Quantifying their composure is a little beyond the scope of this post — it’d require data from their misses too. Another idea would be to look at how close a Pitch Control window is to being closed down, before the ball is passed off (timing of the pass).
But you cannot just sit back all game
But, the answer isn’t to sit back all game. Inviting Liverpool to take on their their attacking shape has proved to be futile too. Take a look at the PPDA (Passes Per Defensive Action) that Liverpool were allowed by various teams during 2018–20. The more passes you allow a team to make per defensive action, the less intense/aggressive your pressing.
3: Watch out for the Press
Klopp’s pressing style was once notorious for its intensity. Liverpool, now with a few years of experience under this system, seem to have mastered the conservative press: maximising their outcome for limited energy expenditure.†
Liverpool crowd the ball with 3 players and the touchline to win the ball in the Chelsea half. A Chelsea defender gets dragged out by Firmino, who is able to cut-back to a wide open Mane due to the very defender he dragged away.
Firmino is a very active member of the Liverpool press, and often occupies highly unnatural positions for a striker during this. Other times, he drops deep during build up which tasks Salah and Mane with making penetrating diagonal runs into the box. Every defender has to adhere to his zonal responsibility, otherwise the entire defence is prone to getting sheared apart due to this unpredictability.
In this dataset, we see 2 teams with their modern take on build-up play: which employs 3-at-the-back, with fullbacks providing ample attacking width to feed a narrower attacking front. This approach is simply a liability against Liverpool’s pressing; as a back-3 is usually not equipped to handle Liverpool’s precise runs once they win the ball.
4: Fluidly Shift the Defence Laterally
- Once Liverpool build into their attacking formation, the defence must be extremely fluid in responding to switches of play. This will mean passing off your man responsibility to the adjacent defender/midfielder.
- In practice, this requires an extremely well-drilled defence that will move as a unit, mirroring whichever side of the pitch Liverpool is attacking from (otherwise gaps will open through the middle). Refer to Atletico’s defensive organisation in Anfield 19/20.
5: Keep the Fullbacks Honest
As we’ve already seen, just defending with 11 men is simply not an option. Having attackers occupy the spaces Trent and Robertson leave behind is important to stop them building a full head of steam.
Of course, these attackers will not be of much use over there without any service. The solution is to imitate Liverpool’s own direct passing, feeding long balls to the wide attackers who can keep the Liverpool fullbacks busy. Of course we do not have concrete examples of this working from the dataset; but we refer to the games against Watford in 18/19, and Napoli, Atletico in 19/20.
6: Frustrate Them
In a world of maths and analytics, this will come across as lax and non-rigorous. But we all know it: football is a human game. There’s a psychological component to the game which our models don’t yet encapsulate. Crowd noise, game rhythm, physicality, tactical fouls, refereeing — all have an effect on players.
With each Liverpool player capable of sheer individual brilliance at a moment’s notice (see Mane and TAA goals above), getting Liverpool to play below their xG is sometimes the only option. Sometimes that’s just down to luck. Look no further than the Atletico knockout, where the match finished 2–3 to Liverpool, but had an xG of 2.9–1.1!
Appendix 1: Pass Difficulty
Metrics such as xT and VAEP are probabilistic, they estimate pass probabilities on empirical data; which could generalise beyond a Physics-based model asymptotically. So the idea here is to bootstrap ourself with known information about making passes — the physics. A big motivating factor is that players only make passes they are comfortable with — which skews the statistical insight we could provide on relative passing abilities of players by just considering passes attempted in games.
The pass physics model from Spearman et. al considers the reception of the pass — i.e. which team is likely to control the ball given a ball lands there. So here, we propose to also capture the variance resulting from an intent to pass the ball to a certain location. Easy passes would have minimum variance, but some of the crosses by Trent would have a higher variance/difficulty. Similar to Spearman et. al, each opponent that has a chance to intercept decreases the probability of the original pass reaching its target, i.e. increasing its difficulty.
Using simple geometry, we can hypothesise the features that build a PD model: the permitted pass angles such that it reaches the target player. Permitted pass elevation trajectories (close by defenders would require you to chip and dip the ball accurately. Tight allowances on any of elevation, spin, angle would significantly increase the difficulty. To put it in perspective, Trent’s cross against Norwich flies past around 4 defenders, with very little room for error in elevation.
Calculating PD ratings as part of player profiles could then be used to shift the PC surface to reflect the abilities (and tendencies) of the player in possession to attempt more difficult passes. This would be seen as expanding the PC surface of teams when creative midfielders are in possession. This could open the door to new attacking and defensive strategies. Training a regressor against match data, which minimises the error between our model’s PD scores and some suitable inverse of pass success (from data) would then give us the relative weights of the features. This would need more time, and a bigger dataset to implement — beyond the scope of this post. A very interesting possibility nonetheless.
Appendix 2: Liverpool Pressing
We look at PPDA to quantify Liverpool’s press. PPDA between 4–8 is usually indicative of aggressive pursuit of the ball, i.e. as high press. But over the past 3 seasons we see this interesting trend: Liverpool have been progressively lowering the intensity of their press, 16/17 (8.05), 17/18 (9.45), 18/19 (10.08). It appears that they are now much more conservative with their press, and have the experience to turn it up as and when needed. More from Tifo Football.
Code + Video:
Update: This article was selected as one of the three winning entries to the Friends of Tracking Liverpool Analytics challenge!
FoT graciously invited us to present our entries on one of the live YouTube sessions: https://www.youtube.com/watch?v=AFm3JNPu9Jw&t=8m16s
And code for the visualisations on Github: https://github.com/suryako/FoT-Liverpool-Analytics-Surya
Acknowledgements
Many thanks to David Sumpter and Friends of Tracking for organising this, Ricardo Tavares for providing the dataset, and Laurie Shaw for the visualisation library!
